Warn user if it looks like they chose the wrong positive class (roc_auc < 0.5)

[Aylr]

With the advent of guess_positive_label(), and the ability for a user to specify binary_positive_label when instantiating TrainedSupervisedModel, it's possible that they selected the incorrect class. This manifests as an inverted ROC cure and a ROC AUC < 0.5.

This should be easily detectable and we could warn the user on the plot and anytime metrics are accessed.